{
  "id": 231060,
  "title": "The more labeled images, the better score",
  "url": "/competitions/hubmap-kidney-segmentation/discussion/231060",
  "author_name": "Swikwislkdjc",
  "post_date": "2021-04-06T19:22:12.182000",
  "votes": 9,
  "comment_count": 8,
  "views": 0,
  "content": "<p>The paper is from Nature Medicine. The authors use histology images to predict whether the image has tumors or not. As shown in figure A (below), they compared the relationship between the amount of images and the validation error. They found that <strong>when there are more labeled data, the error becomes smaller.</strong> In fact, the error of 1000 whole slide images for training is much smaller than that of 100 whole slide images. <br>\n<img src=\"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7418463/bin/nihms-1609511-f0002.jpg\" alt=\"\"></p>",
  "messages": [
    {
      "id": 1265305,
      "postDate": "2021-04-06T19:22:12.183Z",
      "content": "<p>The paper is from Nature Medicine. The authors use histology images to predict whether the image has tumors or not. As shown in figure A (below), they compared the relationship between the amount of images and the validation error. They found that <strong>when there are more labeled data, the error becomes smaller.</strong> In fact, the error of 1000 whole slide images for training is much smaller than that of 100 whole slide images. <br>\n<img src=\"https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7418463/bin/nihms-1609511-f0002.jpg\" alt=\"\"></p>",
      "rawMarkdown": "The paper is from Nature Medicine. The authors use histology images to predict whether the image has tumors or not. As shown in figure A (below), they compared the relationship between the amount of images and the validation error. They found that **when there are more labeled data, the error becomes smaller.** In fact, the error of 1000 whole slide images for training is much smaller than that of 100 whole slide images. \n![](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7418463/bin/nihms-1609511-f0002.jpg)",
      "votes": 9
    },
    {
      "id": 1271493,
      "postDate": "2021-04-12T16:56:24.790Z",
      "content": "<p>There is a dataset here that correlates to this competition, its a 1024px, so you need to do a little tweaking before implement it together with your model.</p>\n<p><a href=\"https://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024\" target=\"_blank\">https://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024</a></p>\n<p>I will be working on this for the next few days, let me know if it helps you!</p>",
      "rawMarkdown": "There is a dataset here that correlates to this competition, its a 1024px, so you need to do a little tweaking before implement it together with your model.\n\nhttps://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024\n\nI will be working on this for the next few days, let me know if it helps you!",
      "votes": 1,
      "replies": [
        {
          "id": 1271928,
          "postDate": "2021-04-13T04:32:39.340Z",
          "rawMarkdown": "",
          "isDeleted": true
        },
        {
          "id": 1272326,
          "postDate": "2021-04-13T12:03:48.630Z",
          "content": "<p>Best i got without this dataset is 0.9363 on eff2, i still have to implement this dataset, probably will be able to do so by this weekend.</p>\n<p>I have been trying to ensemble models on this kernel and can't see a way to get past 0.915 on LB, did you ensemble several models or are using just one? it seems that using just fewer (or even just ine) is getting me better results… </p>",
          "rawMarkdown": "Best i got without this dataset is 0.9363 on eff2, i still have to implement this dataset, probably will be able to do so by this weekend.\n\nI have been trying to ensemble models on this kernel and can't see a way to get past 0.915 on LB, did you ensemble several models or are using just one? it seems that using just fewer (or even just ine) is getting me better results... ",
          "votes": 1
        },
        {
          "id": 1272552,
          "postDate": "2021-04-13T15:06:20.653Z",
          "rawMarkdown": "",
          "isDeleted": true
        }
      ]
    },
    {
      "id": 1267190,
      "postDate": "2021-04-08T11:19:22.667Z",
      "content": "<p>Good idea!</p>",
      "rawMarkdown": "Good idea!"
    },
    {
      "id": 1265572,
      "postDate": "2021-04-07T03:18:25.013Z",
      "content": "<p>Can you share the link of the paper?</p>",
      "rawMarkdown": "Can you share the link of the paper?",
      "replies": [
        {
          "id": 1267749,
          "postDate": "2021-04-08T18:40:24.910Z",
          "content": "<p><a href=\"https://pubmed.ncbi.nlm.nih.gov/31308507/\" target=\"_blank\">https://pubmed.ncbi.nlm.nih.gov/31308507/</a></p>",
          "rawMarkdown": "https://pubmed.ncbi.nlm.nih.gov/31308507/"
        }
      ]
    },
    {
      "id": 1268007,
      "postDate": "2021-04-09T04:05:32.710Z",
      "rawMarkdown": "",
      "isDeleted": true
    }
  ],
  "comments": [
    {
      "id": 1271493,
      "author_name": "victorasso",
      "author_url": "",
      "post_date": "2021-04-12T16:56:24.790000",
      "content": "<p>There is a dataset here that correlates to this competition, its a 1024px, so you need to do a little tweaking before implement it together with your model.</p>\n<p><a href=\"https://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024\" target=\"_blank\">https://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024</a></p>\n<p>I will be working on this for the next few days, let me know if it helps you!</p>",
      "votes": 1,
      "replies": [
        {
          "id": 1271928,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-04-13T04:32:39.340000",
          "content": "",
          "votes": 0,
          "replies": []
        },
        {
          "id": 1272326,
          "author_name": "victorasso",
          "author_url": "",
          "post_date": "2021-04-13T12:03:48.630000",
          "content": "<p>Best i got without this dataset is 0.9363 on eff2, i still have to implement this dataset, probably will be able to do so by this weekend.</p>\n<p>I have been trying to ensemble models on this kernel and can't see a way to get past 0.915 on LB, did you ensemble several models or are using just one? it seems that using just fewer (or even just ine) is getting me better results… </p>",
          "votes": 1,
          "replies": []
        },
        {
          "id": 1272552,
          "author_name": "",
          "author_url": "",
          "post_date": "2021-04-13T15:06:20.653000",
          "content": "",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1267190,
      "author_name": "ZavodRobotov",
      "author_url": "",
      "post_date": "2021-04-08T11:19:22.667000",
      "content": "<p>Good idea!</p>",
      "votes": 0,
      "replies": []
    },
    {
      "id": 1265572,
      "author_name": "Paceiver",
      "author_url": "",
      "post_date": "2021-04-07T03:18:25.013000",
      "content": "<p>Can you share the link of the paper?</p>",
      "votes": 0,
      "replies": [
        {
          "id": 1267749,
          "author_name": "Swikwislkdjc",
          "author_url": "",
          "post_date": "2021-04-08T18:40:24.910000",
          "content": "<p><a href=\"https://pubmed.ncbi.nlm.nih.gov/31308507/\" target=\"_blank\">https://pubmed.ncbi.nlm.nih.gov/31308507/</a></p>",
          "votes": 0,
          "replies": []
        }
      ]
    },
    {
      "id": 1268007,
      "author_name": "",
      "author_url": "",
      "post_date": "2021-04-09T04:05:32.710000",
      "content": "",
      "votes": 0,
      "replies": []
    }
  ],
  "raw_markdown_by_id": {
    "1265305": "The paper is from Nature Medicine. The authors use histology images to predict whether the image has tumors or not. As shown in figure A (below), they compared the relationship between the amount of images and the validation error. They found that **when there are more labeled data, the error becomes smaller.** In fact, the error of 1000 whole slide images for training is much smaller than that of 100 whole slide images. \n![](https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7418463/bin/nihms-1609511-f0002.jpg)",
    "1271493": "There is a dataset here that correlates to this competition, its a 1024px, so you need to do a little tweaking before implement it together with your model.\n\nhttps://www.kaggle.com/baesiann/glomeruli-hubmap-external-1024x1024\n\nI will be working on this for the next few days, let me know if it helps you!",
    "1267190": "Good idea!",
    "1265572": "Can you share the link of the paper?",
    "1268007": ""
  }
}